Critères de succès pour un agent IA
A success criterion is the explicit definition of what 'done' means for a given task, written before you launch the agent. Most agent failures don't come from its incompetence but from a goal that stayed in your head: you ask it to 'tidy the kitchen' without saying what tidy means to you, so the agent stops at ITS own idea of finished. Anthropic makes this the first step of any Claude application (the 'Define your success criteria' doc) with three demands: be specific, measurable and achievable.
Strengths
- Free, immediate lever: three lines written before launch, no stack change
- Gives the agent a clean stop condition: it stops continuing until it breaks something
- Forces you to clarify your own goal: if you can't make it binary, you didn't know what you wanted
Limitations
- Requires upfront effort many skip: writing the criterion takes more discipline than a vague prompt
- A badly framed binary criterion can over-constrain: too rigid, the agent rejects correct but unforeseen outputs
Best for
- A PM or PO delegating tasks to an agent and systematically getting back work that is "almost done"
- A designer or DS-manager who wants to drive an AI without being an engineer: setting a verifiable criterion is within reach
- A tech lead automating agents continuously who needs reliable stop conditions